Open-source (Apache 2.0) LLM series 'born for reasoning.' Pre-trained & RL-tuned models (like the 7B) match o1-mini on math/code. Base/SFT/RL models released.
This is the 7th launch from MiMo. View more
MiMo-V2.6
Launching today
MiMo-V2.6 is Xiaomi’s open omnimodal model family for long-horizon agent work. Pro and Flash handle text, images, audio and video with 1M context, while Xiaomi is also releasing the technical report, RL environments and training code behind the public post-training run.
Free
Launch Team
Flowtica Scribe
Hi everyone!
MiMo-V2.6 comes in Pro and Flash, both native omni models, plus an UltraSpeed mode when you want the tokens sooner. API pricing stays the same as V2.5.
Early in post-training, they put up a public board and streamed the RL run live to anyone watching. That’s the first model lab I’ve seen do this in the open.
Under six days, you could literally watch the curves move as both models went through 30 RL steps. Pro climbed from 58.4 to 72.6 on DeepSWE, while Flash went from 48.8 to 65.7.
The release itself goes pretty far beyond the checkpoint. Xiaomi is open-sourcing the weights, technical report, RL environments, training framework and the lightweight harnesses they used around it.
Pro now scores highest among open models on Artificial Analysis, while the same family is doing some surprisingly broad work across coding, computer use, 3D and creative tasks.